Enhanced diffusion-based model for rubber stamp removal

Existing models for stamp removal are primarily designed for color stamps, yet lack specific models for removing stamps from photocopied documents. In this work, the first model designed for removing stamps from photocopied documents is proposed. It incorporates two groundbreaking technologies: Markov process-based Attention Mechanism (MAM) and Bionic Rectified Linear Unit (BReLU). MAM enhances model performance by applying Markov process principles to the attention mechanism, enabling a dynamic and iterative adaptation that draws focus precisely where it’s needed. Furthermore, BReLU replicates the complex and varied action potentials of human neurons, enhancing the model’s learning ability and realism. The proposed model achieved superior performance metrics in stamp removal tasks, obtaining a PSNR of 41.62, an SSIM of 0.993, and an RMSE of 3.78. The images generated by the proposed model closely approximate the ground truth. This pioneering work opens new avenues for research and sets the stage for future innovations.